Bulk AI text generation without the frontier-model bill.

This is the lowest-friction 50% saving available if you are already an OpenAI customer: same models, same outputs, one file upload instead of a loop of live calls. Batch requests draw from a separate rate-limit pool, so a large overnight job does not eat the quota your production traffic needs during the day.

Best for: Teams already standardised on OpenAI models that want to halve the bill on overnight jobs without changing model behaviour, prompts, or evaluation baselines.

Caveat: Everything is queued. Unfinished requests in an expired batch are cancelled, and you are still billed for the ones that completed, so oversized single batches are a real risk.

The 2 GB input file ceiling is the standout number here — it is an order of magnitude above what most competitors accept in one submission, which matters when you are regenerating an entire catalogue in one pass. Batch supports the same modalities as the interactive API, so document and image inputs come along for the ride.

Best for: Very large corpora, mixed text and document workloads, and teams that want a hyperscaler contract behind a bulk pipeline.

Caveat: Some listed prices are explicitly promotional. Gemini 3.7 Flash pricing doubles on 1 January 2027 per Google's own pricing page, so model the post-promotion rate before committing a year of volume.

The practical draw is the batch ceiling and the turnaround. A single batch takes up to 100,000 requests or 256 MB, and Anthropic states most batches finish in under an hour rather than using the full 24-hour window — which makes it usable for same-day content operations, not only overnight jobs. Results stay retrievable for 29 days from creation.

Best for: Long-form drafting, editorial rewriting, summarisation of sensitive material, and bulk jobs where a weaker model would cost you more in human editing than it saves in tokens.

Caveat: Claude models are not the cheapest per token even after the discount, so this only wins when output quality reduces downstream editing. The 29-day clock starts at batch creation, not completion.

DeepSeek's structure is unusual and useful for bulk work: instead of trading latency for a batch discount, it discounts by clock time, so a scheduled overnight run gets the cheaper rate while staying synchronous. Cache-hit input pricing is low enough that repeated system prompts across thousands of calls stop mattering.

Best for: Cost-driven bulk generation where you can schedule the run into the cheaper window and evaluate quality prompt family by prompt family.

Caveat: Peak and off-peak windows are defined in UTC and have changed before, so re-read the pricing page before scheduling a pipeline around them. Run your own quality and data-policy review as you would for any provider.

Together's value in bulk work is optionality: the same batch interface fans out across a broad open-model catalogue, so you can move a workload to a cheaper or better-fitting model without rewriting the pipeline. Batch rate limits are independent of your real-time usage.

Best for: Teams that do not want to bet the pipeline on one closed vendor, and workloads where an open model is good enough for the task at hand.

Caveat: The 24-hour window is best effort rather than a guarantee, and cancelling a running job still bills the responses already completed.

Groq is the only provider here that lets you set a completion window of up to seven days rather than the industry-standard 24 hours. For a genuinely large job that is a feature: a longer window gives the scheduler more room, and you are not forced to shard the work into daily chunks to fit a deadline you do not actually have.

Best for: Very large one-off generation runs — a full site rebuild, a back-catalogue rewrite — where a longer, more relaxed window buys you better scheduling.

Caveat: Batch covers chat completions and audio endpoints only, and the discount explicitly does not stack with caching, so heavily cached prompts may be cheaper synchronously.

Two numbers set Fireworks apart. Input datasets go up to 80 GiB, which removes sharding from the problem entirely for most teams. And unlike Groq, the batch and prompt-caching discounts stack, so a workload with a long fixed instruction block and short variable inputs — exactly the shape of bulk SEO and product-description generation — gets the deepest discount available here.

Best for: The largest bulk runs on this list, and pipelines with a heavy shared system prompt where the caching discount compounds across every request.

Caveat: There is no guaranteed turnaround: jobs sit in a pending state and may pause while waiting on capacity. Pick the expiry window deliberately rather than defaulting.

The one-million-request ceiling is twenty times OpenAI's and ten times Anthropic's, which makes Mistral the natural choice when the unit of work is enormous and uniform. Batch covers an unusually wide endpoint list including embeddings, moderation, classification, OCR, and transcription — useful when bulk generation is one stage of a longer pipeline.

Best for: European teams with data-residency requirements, multilingual bulk generation, and jobs large enough that a 50,000-request cap would force sharding.

Caveat: Mistral does not publish a standard completion window, only a TIMEOUT_EXCEEDED job status, so you cannot plan a deadline against a documented SLO the way you can with OpenAI or Google.

OpenRouter load-balances by price by default, weighting selection inversely to cost squared, and will fail over to other providers during an outage. For long-running bulk jobs that resilience is worth more than the last few cents per million tokens. You can sort by throughput instead of price, set minimum-throughput and maximum-latency floors at p50 to p99, and restrict routing to providers that do not collect data or that enforce zero data retention.

Best for: Bulk pipelines that must not stop when one provider degrades, and teams still deciding which model to standardise on.

Caveat: It is a layer, not a discount. You still own the monitoring: which route actually served each request, what it cost, and whether its data policy matches your obligations.